AN IMPROVED ILLUMINATION NORMALIZATION

APPROACH BASED ON WAVELET TRANFORM FOR

FACE RECOGNITION FROM SINGLE TRAINING

IMAGE PER PERSON

Chun-Nian Fan

1, 2

and Fu-Yan Zhang

Department of Computer Science and Technology, Nanjing University, Nanjing, China

Computer and Software Institute, Nanjing University of Information Science & Technology, Nanjing, China

Keywords: Face Recognition, Wavelet Analysis, Illumination Normalization, Histogram Equalization.

Abstract: Recent research on face recognition shows that the illumination change is one of the key issues remaining to

be addressed. To recognize faces under varying illuminations with single training image per person

conditions, we propose an improved wavelet-based normalization method. We use wavelet transform to

decompose an image into its low frequency and high frequency components. Then, we apply histogram

equalization to the low frequency coefficients and de-noise the high frequency coefficients adaptively.

Lastly, the high frequency coefficients are accentuated by multiplying by a scalar so as to enhance edges. A

normalized image is obtained from the modified coefficients by inverse wavelet transform. Among others,

the proposed method has the following advantages: (1) it does not need any prior information of 3D shape

or light sources, and it aims at addressing illumination issue for face recognition from only one training

image per person; (2) due to the multiscale nature of wavelet transform, it has better edge-preserving ability

in low frequency illumination fields; and (3) it is computationally feasible and fast. We use PCA method to

recognize normalized image with only one training image. The experimental results obtained by testing on

the Yale face database B demonstrate the effectiveness of our method with significant improvement in the

face recognition system.

1 INTRODUCTION

Related research on face recognition in recent years

has made great progress (Zhao et al., 2003), a

number of FR systems achieved good performance

in the latest report of Face Recognition Vendor Test

(FRVT 2006), yet many issues still remain to be

addressed. Illumination changes are one of the key

issues (Adini et al., 1997). To deal with the

illumination variation problem, many methods have

been proposed, which can be roughly classified into

three categories. (1) illumination normalization:

these methods use image processing techniques such

as histogram equalization (HE), Gamma correction,

and logarithmic transformation (Shan et al., 2003;

Savvides and Kumar, 2003), to normalize human

face image in order to obtain face image’s stability

under illumination changes. However, these methods

have limited success in handling arbitrary

illumination changes. (2)invariant feature extraction:

this approach attempts to extract facial features

which are invariant to illumination variations, such

as edge maps, derivatives of the gray-level,

Gabor-like filters methods, and wavelet-based

method (Adini et al., 1997; Zhang et al., 2009).

Empirical studies, however, show that none of these

representations are sufficient to overcome image

variations due to changes in the direction of

illumination. (3)face modeling: some researchers

attempt to construct a generative 3-D face model that

can be used to render face images with different

poses and under varying lighting conditions . A

generative model called illumination cone is

presented in (Belhumeur et al., 1998; Georghiades

and Belhumeur, 2001). One of the drawbacks of the

model-based approaches is that a number of images

of the same object under varying lighting conditions

or 3-D shape information are needed during the

training phase. This drawback limits its applications

in practical face recognition systems. In addition,

473

Fan C. and Zhang F. (2010).

AN IMPROVED ILLUMINATION NORMALIZATION APPROACH BASED ON WAVELET TRANFORM FOR FACE RECOGNITION FROM SINGLE

TRAINING IMAGE PER PERSON.

In Proceedings of the International Conference on Computer Vision Theory and Applications, pages 473-476

DOI: 10.5220/0002785304730476

 SciTePress

existing model-based approaches assume that the

human face is a convex object, i.e. the casting

shadows are not considered. The specularity problem

is also ignored even though the human face is not a

perfect Lambertian surface.

Du and Ward present a wavelet-based

illumination normalization method(Du and Ward,

2005). In their work, an image is decomposed into its

low frequency and high frequency components, then

apply histogram equalization to the approximation

coefficients and at the same time accentuate the detail

coefficients by multiplying by a scalar so as to

enhance edges. A normalized image is obtained from

the modified coefficients by inverse wavelet

transform. However, the obtained high frequency

components using wavelet transform are mixed with

noise, if we directly enlarge the high frequency

components, the noise will also be enlarged, and that

will hinder the face recognition process. Therefore, if

we de-noise the high frequency components before

accentuating, that will remove noise and enhance the

edges. In addition, Du and Ward select 1-level

wavelet decomposition, which is not the optimal

decomposition level. In this paper, an improved

wavelet-based illumination normalization method is

proposed. Firstly, we decompose an image into its

low frequency and high frequency components using

3-level wavelet decomposition. Secondly, we apply

histogram equalization to the approximation (low

frequency) coefficients. Thirdly, we de-noise the

detail (high frequency) coefficients adaptively and

accentuate them by multiplying a scalar so as to

enhance edges. Lastly, a normalized image is

obtained from the modified coefficients by inverse

wavelet transform. The resulted image not only

enhances contrast, but also enhances edges and

details that will facilitate the further face recognition

task. The proposed illumination normalization

method do not need many traning images per person

and then can be apply to face recognition from only

one training image per person conditions. That will

benefit the practical face recognition system since the

current face recognition techniques lies in the

difficulties of collecting samples (Tan et al., 2006).

The remainder of this paper is organized as

follows: Section 2 describes the proposed approach in

detail; Sections 3 and 4 show the experimental results

and our conclusions.

2 THE PROPOSED METHOD

The illumination changes mainly affect the low

frequency components, and the high frequency

components represent detail information such as the

location and shape information of the key facial

features which is essential to further face recognition

(Wei, 2006). Therefore, in this work, we firstly

decompose the face image use wavelet transform.

Then

different band coefficients are manipulated

separately. Lastly, we reconstruct the face image

using inverse wavelet transform. The resulting image

will not only contain the most essential information

for pattern recognition, but also greatly reduce the

influence of illumination changes.

2.1 Wavelet Decomposition

Wavelet transform is a representation of a signal in

terms of a set of basis functions, which is obtained

by dilation and translation of a basis wavelet. Since

wavelets are short-time oscillatory functions having

finite support length (limited duration both in time

and frequency), they are localized in both time

(spatial) and frequency domains. The joint

spatial-frequency resolution obtained by wavelet

transform makes it a good candidate for the

extraction of details as well as approximations of the

images (Li, 2003). For 2D discrete wavelet

transform (DWT), an image is represented in terms

of translations and dilations of a scaling function and

a wavelet function. The scaling and wavelet

coefficients can be easily computed using a 2D filter

bank consisting of low-pass and high-pass

filters.According to (Feng et al., 2000), we choose

3-level db1 wavelets in our work. After applying a

3-level of 2D decomposition, an image is

decomposed into subbands of different frequency

components as shown in Figure 1 and Figure 2.

Figure 1: Multi-resolution

structure of wavelet

decomposition of an image.

Figure 2: Wavelet

decomposition of a face

image.

2.2 Histogram Equalization of the

Approximation Coefficients

The histogram of an image is usually used to

determine which particular gray scale transformation

is required to enhance the image contrast. Histogram

VISAPP 2010 - International Conference on Computer Vision Theory and Applications

474

equalization (HE) is one of the most useful contrast

enhancement schemes. When an image’s histogram

is equalized, image pixel values are mapped to

uniformly distributed pixel values, as much as

possible. In this paper, we apply histogram

equalization to the approximation coefficients, after

that, the illumination of the approximation image is

also normalized.

2.3 De-noising and Highlighting the

Detail Coefficients

There are two thresholding schemes for general

wavelet-based image de-noising: hard thresholding

and soft thresholding (Donoho, 1995). Soft

thresholding eliminates the discontinuity (at the

threshold) that is inherent in hard thresholding. In

this paper, we employ a soft thresholding operation

on the detail coefficients. The de-noising procedure

removes the "noise” signal by thresholding only the

wavelet coefficients of the detail subbands, while

keeping the low resolution coefficients unaltered.

Thus, it is able to keep sharp edges information in

low frequency illumination fields.

After de-noising, we multiply each element in the

detail coefficient matrix with a scale factor (>1) to

enhance edges. Different scale factors will lead to

different results. Figure 3 shows the face recognition

rates using different scale factor, and we can see that

the optimal scale factor is near 3.

69.0%

69.2%

69.4%

69.6%

69.8%

70.0%

70.2%

70.4%

70.6%

70.8%

71.0%

12345

scale factor

average Top1 recognition rate(%)

Figure 3: Face recognition rate using different scale factor.

（a） (b) (c) (d)

Figure 4: Reconstructed images comparison (a) original;

(b) histogram equalized; （ c ） wavelet-based method

normalized; (d) the proposed method normalized.

2.4 Image Reconstruction

The enhanced image is reconstructed from the

histogram equalized approximation coefficients and

the enlarged detail coefficients in all three directions

using inverse wavelet transform. The results of

applying the proposed method on two different

images of two different persons are shown in Figure

4. The enhanced images using the proposed method

are sharper and have more details and more suitable

for face recognition intuitively.

3 EXPERIMENTS

To evaluate the performance of the proposed

illumination normalization method, we test it on the

Yale face database B (Lee et al., 2005). The Yale

face database B contains 5760 single light source

images of 10 subjects. Each subject has 9 poses and

each pose has 64 different illumination conditions.

Since this paper mainly deals with the illumination

problem, we only choose the 64 frontal pose images

captured under 64 different lighting conditions for

each of the ten persons. The images are divided into

five subsets according to the light-source directions

(azimuth and elevation): Subset 1 (angle < 12

degrees from optical axis), Subset 2 (20 < angle < 25

degrees), Subset 3 (35 < angle < 50 degrees), Subset

4 (60 < angle < 77 degrees), and Subset 5 (others).

26.0%

36.0%

46.0%

56.0%

66.0%

76.0%

86.0%

96.0%

subset1 subset2 subset3 subset4 subset5 average

Top1 recognition rate(%)

Ours

Figure 5: Top 1 recognition rate comparisons.

In our experiment, we directly use PCA method

for face recognition with only single training sample.

The face image with normal illumination in subset 1

(only one image for each person) is chosen as the

gallery and each of the images in the 5 subsets is

matched to the images in the gallery so as to find a

best match. All test image data are manually aligned,

cropped, and re-sized to 168x192 images. The

recognition rates use the Euclidean distance

nearest-neighbor classifier. We compare the face

AN IMPROVED ILLUMINATION NORMALIZATION APPROACH BASED ON WAVELET TRANFORM FOR

FACE RECOGNITION FROM SINGLE TRAINING IMAGE PER PERSON

475

recognition performance of different methods

including histogram equalization, the wavelet-based

method (Du and Ward, 2005), and our method.

Conforming to the FERET test rules (Phillips et al.,

1998), we have not only tested the Top 1 recognition

rate, but also tested the Top 3 recognition rate.

The recognition rates are illustrated in Figure 5

and Figure 6. It is shown from Figure 5 and Figure 6

that our proposed method outperforms the histogram

equalization method and the wavelet-based method

at every single subset.

60.0%

65.0%

70.0%

75.0%

80.0%

85.0%

90.0%

95.0%

100.0%

subset1 subset2 subset3 subset4 subset5 average

Top3 recognition rate(%)

Ours

Figure 6: Top 3 recognition rate comparisons.

4 CONCLUSIONS

This paper presents an improved wavelet-based

illumination normalization method for face

recognition from only one training image per person.

The proposed approach has not only enhanced

contrast, but also enhanced edges and details that will

facilitate the further face recognition task. There is no

need to any prior information of 3D shape and light

sources. Moreover, due to the multiscale nature of

wavelet transform, it has better edge-preserving

ability in low frequency illumination fields. In

addition, it is computationally feasible and fast. The

experimental results obtained by testing on the Yale

face database B demonstrate the effectiveness of our

method with significant improvement in the face

recognition system.

REFERENCES

Adini, Yael, Yael Moses and Shimon Ullman. (1997).

Face Recognition: The Problem of Compensating for

Changes in Illumination Direction. IEEE Transactions

on Pattern Analysis and Machine Intelligence, 19(7),

721-732.

Belhumeur, P. N. and D. J. Kriegman. (1998). What is the

set of images of an object under all possible

illumination conditions. International Journal of

Computer Vision, 28(3), 245–260.

Donoho, D. L. (1995). De-noising by Soft-thresholding.

IEEE Transaction on Information Theory, 41(3),

613-627.

Du, Shan and Rabab Ward. (2005). Wavelet-based

illumination normalization for face recognition. In

IEEE International Conference on Image

Processing(ICIP 2005).

Feng, G. C., P. C. Yuen and D.Q. Dai (2000). Human face

recognition using PCA on wavelet subband. Journal of

Electronic Imaging, 2(9), 226-233.

Georghiades, Athinodoros S. and Peter N. Belhumeur.

(2001). From Few to Many Illumination Cone Models

for Face Recognition Under Variable Lighting and

Pose. IEEE Transactions On Pattern Analysis And

Machine Intelligence, 23(6), 643-660.

Lee, K.-C.J., Ho, and D. J. Kriegman. (2005). Acquiring

linear subspaces for face recognition under variable

lighting. IEEE Transactions On Pattern Analysis And

Machine Intelligence, 27(5), 684–698.

Li Bicheng，Luo Jianshu. (2003). Wavelet Analysis and

Its Application. Beijing ： Publishing House of

Electronics Industry.

Phillips P. J., et al. (2007) FRVT 2006 and ICE 2006

Large-Scale Results [Online]. Available:

http://www.frvt.org/FRVT2006/

Phillips P. J., Wechsler H., Huang J.and Rauss P. J.

(1998). The FERET database and evaluation procedure

for face recognition algorithms. Image and Vision

Computing, 16(5), 295-306.

Savvides ， M. and V. Kumar. (2003). Illumination

normalization using logarithm transforms for face

authentication. In Proc. IAPR AVBPA, 549–556.

Shan, Shiguang, Wen Gao, Bo Cao and Debin Zhao.

(2003). Illumination normalization for robust face

recognition against varying lighting conditions. In

Proceedings of the IEEE International Workshop on

Analysis and Modeling of Faces and Gestures

(AMFG’03).

Tan, X., Chen, S., Zhou, Z.-H. & Zhang, F. (2006) Face

recognition from a single image per person: A survey.

Pattern Recognition, 39, 1725-1745.

Wei Wang, Jiatao Song, Zhongxiu Yang, Zheru Chi.

(2006). Wavelet-based Illumination Compensation for

Face Recognition using Eigenface Method. In

Proceedings of the 6th World Congress on Intelligent

Control and Automation, June 21 - 23, Dalian,

China.

Zhang, Taiping, Bin Fang, Yuan Yuan, Yuan Yan Tang,

Zhaowei Shang, Donghui Li and Fangnian Lang.

(2009). Multiscale facial structure representation for

face recognition under varying illumination. Pattern

Recognition

42, 251-258.

Zhao, WenYi, R. CHELLAPPA, P. J. PHILLIPS and A.

ROSENFELD. (2003). Face Recognition: A Literature

Survey. ACM Computing Surveys, 35(4), 399–458.

VISAPP 2010 - International Conference on Computer Vision Theory and Applications

476